slides - fasig - photon 12 - durham - september 2012

15
1 Challenge the future TU Delft A Machine Learning Approach to Fringe-Location Identification Firas Sawaf and Roger Groves FASIG - Photon 12 Durham September 2012

Upload: tudelft

Post on 05-Nov-2023

0 views

Category:

Documents


0 download

TRANSCRIPT

1Challenge the future

TU DelftA Machine Learning Approachto Fringe-Location IdentificationFiras Sawaf and Roger Groves ● FASIG - Photon 12 ● Durham ● September 2012

2Challenge the future

Fringe LocationEdge detection

3Challenge the future

Fringe LocationPre-filtering

4Challenge the future

A Machine Learning ApproachFree online course

5Challenge the future

A Machine Learning ApproachNeural networks

6Challenge the future

A Machine Learning ApproachTraining

7Challenge the future

A Machine Learning ApproachTesting, training set 99.9%

8Challenge the future

A Machine Learning ApproachTesting, over-fitting

9Challenge the future

A Machine Learning ApproachTesting, cross-validation set 89.9%

10Challenge the future

A Machine Learning ApproachTesting, cross-validation set

MCC% = Threshold% * MCC

11Challenge the future

A Machine Learning ApproachIntuition

• AA: Ability to Absorb

• AG: Ability to Generalise

• AIM: Absorption vs. Ideal Measure

• SAG: Spread of Absorption vs. Generalisation

12Challenge the future

A Machine Learning ApproachPerformance - Large AIM, Small SAG

120,000 training examples, 200 nodes in hidden layer

13Challenge the future

A Machine Learning ApproachPerformance - Medium AIM, Large SAG

30,000 training examples, 800 nodes in hidden layer

14Challenge the future

A Machine Learning ApproachPerformance - Small AIM, Large SAG

30,000 training examples, 1600 nodes in hidden layer

15Challenge the future

A Machine Learning ApproachPerformance - Medium AIM, Medium SAG

120,000 training examples, 1600 nodes in hidden layer